Sensing signals

We’ve been talking about AI for 70 years. Specifically, since the summer of 1956, when a small group of researchers gathered in Hanover, New Hampshire, to brainstorm about the possibility that machines could simulate every aspect of human intelligence. It was a founding moment, not only because it was there that John McCarthy coined the term “artificial intelligence,” but more importantly because it defined the academic relevance of AI as a field of research and set the stage for attracting funding for the research that would prove decisive decades later (though the participants had envisioned a more rapid impact). There were only about a dozen participants, but they were the great visionaries of their fields. In addition to McCarthy, considered the father of AI, the conference included Herbert Simon (Nobel Prize in Economics in ’78) and Allen Newell, who together presented the Logic Theorist. This program simulated human problem-solving skills and is considered the first artificial intelligence program.

Fast forward 20 years, and we find a group of researchers at Stanford University developing MYCIN, one of the first and most relevant expert systems. Their system, based on some 600 rules, could query physicians with simple, logically linked questions to assess confidence, identify bacterial infections, and suggest antibiotic treatment.

It is a classic example of an expert system, a genre that was gaining momentum at this point in history, a program that uses artificial intelligence techniques to solve problems in a given domain that would normally require the use of human expertise.

Another leap forward 20 years brings us to a defining moment in history. Chess was thought to be so computationally complex that it would be impossible for a machine to solve. Instead, in 1997, in a challenge that has gone down in history, Deep Blue, a supercomputer developed by IBM, proved that a machine, thanks in part to brute force computational supercapacity, could beat Garry Kasparov, the world chess champion at the time. Deep Blue’s victory proved just how powerful the combination of specialized hardware and software optimized for a specific task could be.

Fast forward another 15 years, and we’re at the University of Toronto, where a few students, led by their professors, are developing a new architecture that finally unlocks the potential of neural networks by making them much deeper and more scalable. AlexNet, in particular, was designed to take full advantage of GPU parallelization to speed up the learning phase. It was the dawn of deep learning.

A few years later, building on this concept and improving it with reinforcement learning techniques that allow the computer to play against itself to make the learning phase of the networks even faster and more scalable, DeepMind, a company acquired by Google, developed AlphaGo, a program designed to play the board game Go. Go is much more complex than chess. The number of legal board positions in Go has been calculated to be approximately 2.1×10¹⁷⁰, which is far greater than the number of atoms in the observable universe, which is estimated to be on the order of 10⁸⁰.

In 2016, AlphaGo challenged the world’s strongest Go player, Lee Sedol of Korea. And against all odds, it won. There is a wonderful movie on YouTube that I recommend you watch about the story.

The following year a group of Google researchers published a paper titled Attention Is All You Need, introducing the world to transformer models, which mark a paradigm shift in how machines understand and generate human language.

This chart compares the number of days it took to reach the first million and the first 100 million users. For the first million, Netflix took three years, while ChatGPT took two days. For the first 100 million, Netflix took ten years, while ChatGPT took just over two months. The signal this time became obvious to everybody. Yet signals had been coming in for years. Except that most of them were receivable and decodable only by insiders.

On the one hand, they demonstrate superiority in scaling models further, and on the other, they exhibit generative capabilities at a level not previously achieved. The “T” in GPT (Generative Pre-trained Transformer) stands for transformer.

Indeed, shortly thereafter, OpenAI released GPT1, 2, and 3 sequentially over a 3-year period. Each model was significantly deeper and more powerful than the previous one. GPT-3 was released in 2020. The model has 175 billion parameters, ten times more than GPT-2. Interest grew as several developers connected to the OpenAI API to explore and experiment. But so far, most ordinary people still had marginal exposure or were completely unaware of what was coming.

Until November 2022, when a conversational interface was added to the model, allowing anyone to try talking to ChatGPT. That changed everything. Suddenly, everyone realized how intelligent AI was and how incredible of a leap it had made. And the world went crazy.

This chart compares the number of days it took to reach the first million and the first 100 million users. For the first million, Netflix took three years, while ChatGPT took two days. For the first 100 million, Netflix took ten years, while ChatGPT took just over two months. The signal this time became obvious to everybody. Yet signals had been coming in for years. Except that most of them were receivable and decodable only by insiders.

After almost 70 years of transmission on artificial intelligence, from the founding fathers at the Dartmouth conference and on to MYCIN, Deep Blue, AlphaGo, OpenAI to ChatGPT, the signal finally became visible. We can try to line up these signals as if they were a technology moving toward us, changing in size as they approach. Their size expresses the perceived relevance of that technology to the impact we think it might have.

Not all technologies move the same way; some can be faster than others and grow, some remain constant, some disappear, and some seem promising only to end up deflating and disappointing expectations. It is a game where you have to be good at catching the right signals at the right time. Not easy.

Also because at the same time, there are different technologies appearing and disappearing, growing and decreasing continuously, and approaching at different speeds. How can you put yourself in the best position to seize these opportunities for innovation?

In fact, 80% of executives rank innovation among their top 3 priorities. It is equally clear that something is not working because while it is extremely important, only 10% are satisfied with what is being done.

But what can an organization do to manage innovation? Many things, of course, but I am interested in exploring some simple and affordable ones to understand how tinkering with a few levers can help organizations put themselves in a position to catch the signals and move in time to intercept new technologies and seize the opportunities they present.

The first lever has to do with multidisciplinarity. Catching a signal on a topic requires being tuned in to that frequency. And some frequencies are weak because they broadcast from afar. However, intercepting them in time allows one to arrive ready for impact and grasp the transformative factor of that innovation.

In contrast, verticality in terms of competencies makes it more difficult to catch weak and distant signals. Here, for example, a vertical organization (thus with only one “red” competency) can pick up the signal sent out by a new red technology and move in time to impact it, but having no blue competency, it ignores the signal that passes by without even noticing it.

Here, on the other hand, the organization has a mix of two competencies (red and blue), which allows it to intercept both signals in time and move to impact both technologies. On the other hand, the second organizational lever has to do with the investment logic.

Let us imagine two organizations, with the first having a centralized investment logic and the second a distributed one. Let us also imagine that the two organizations at any given time read an equal mix of signals, and they then intercept and give equal weight to the different signals from the different incoming technologies. The empty isograms represent the different signal levels, which are identical for the two companies.

The company with a centralized strategy tends to overvalue what it perceives to be most relevant and to invest more purposefully on the technology it perceives to be most relevant.

Conversely, the company with a distributed strategy has several decision-making poles, including peripherals, that tend to value weak signals more, even at the expense of the strongest signal. In this case, the colored bars represent investments.

This is an intriguing assumption, but it would be interesting to see how these strategies actually play out in a hypothetical ecosystem. And it may be because when I played SimCity as a kid, I always thought that before a person ran for mayor it would be useful to get them to play the game. I thought a simulation game would be helpful to see these dynamics in action.

And that is what I did, with outside help, by designing a simple simulation game.

Let’s look at the first case here: a company that starts with a single competency (red) and invests everything with a centralized strategy. As a result, it has difficulty recognizing the signals of emerging technologies of different colors. It doesn’t have the expertise to read them, so it fumbles a bit. And because it invests in the center, where it is strong, it struggles to acquire the skills it does not have that it would need to pick up other signals. It scores very low.

The second case is of an entity that again invests centrally, but starts out multidisciplinary. In fact, it starts out better but quickly loses this advantage because it tends to go all in on the biggest signals, leaving little room for exploration.

The third case describes an organization that starts vertically, with a single competence, but adopts a distributed investment logic. So it starts out like the first company did and in fact struggles to intercept technologies because it is a bit blind and does not read the signals well. However, when it finally, even by chance, ends up impacting a technology, it invests in a way that diversifies because it does so with distributed logic. In this case, the initial fatigue is overcome, and having reached a certain level of multidisciplinarity, the company manages to intercept signals well and impact incoming technologies well.

Finally, the last case is of a company that starts out multidisciplinary and continues to invest in a distributed way. The company moves well right away, reads the signals immediately, and reinvests in exploration. It does, in fact, score high.

In fact, trying to play a thousand matches with different combinations of technologies for each strategy combination, we see that the average score is low with the vertical/centralized combination and increases if either of the levers changes. The average is absolutely higher if the multidisciplinary start and distributed investment capacity are combined.

Indeed, the chart shows the average score (white dot) and scoring frequency (violin width). Not only is the mean lower in the former case and higher in the latter, but the probability of occurrence is also centered around the mean.

Obviously, this simple game has no scientific value. However, seeing abstract concepts in action and how they play out under certain conditions is fascinating. Nevertheless, it is valuable food for thought and, in the end, it confirms what seems to me to be strong evidence: in order to maintain a capacity for innovation, organizations must have and keep alive a certain degree of multidisciplinarity, and an extremely effective lever for this is to ensure greater autonomy for the periphery through a decentralized investment strategy.

This seems super evident, obvious to everyone. But it isn’t. At least not for everyone equally.

If you think about it, there are no laws of physics that allow one company to do something that another company cannot do. They are all operating in the same marketplace of capital and talent. And they are all competing for the same potential customers. Except some succeed in creating the conditions that eventually allow them to identify opportunities and be ready to seize them with the right skills at the right time. And others fail to do so.

And for all but a few, the challenge is not with Google or OpenAI. The challenge is with other companies operating in the same context. And if some companies today are wondering which way to go to seize the opportunities or avert the threats of generative artificial intelligence, when in fact some direct competitors have already been experimenting for years or months and have ideas that they are moving forward with, then what is so obvious is actually not.